CAPO Puts a Gate in Front of Protein-Model Fine-Tuning
Protein foundation models have changed how researchers work with sequence and structure, but the bottleneck has shifted from access to capacity. A pretrained model can look impressive in a benchmark and still fail when it meets a specific laboratory assay, a narrower species set, or a time-shifted variant dataset. CAPO, short for Compute-Aware Automated Protein-Model Optimization, addresses that gap by making adaptation a decision problem rather than an automatic reflex.
The central idea is simple and unusually consequential: do not spend GPU cycles until the system has checked whether fine-tuning is feasible, affordable, and worth doing on the target hardware. In this preprint, CAPO is built as a multi-agent system that measures feasibility on the actual accelerator, estimates cost against a fixed budget, runs a canary inside the training loop, and keeps an auditable record of the process. If the requested adaptation is too expensive, CAPO can decline it and pivot to a cheaper route.
That constraint changes the workflow from “train first, explain later” to a controlled admission process for computation.
Measured Admission Control Replaces Blind Training
The practical problem is easy to recognize in protein science. Large pretrained models such as ESM, AlphaFold, and Boltz have broad priors, but those priors are not always aligned with a local assay or downstream benchmark. Fine-tuning can repair that mismatch, yet it also consumes real compute and can fail for reasons that become visible only after hours of work. CAPO treats those failures as avoidable waste.
The system’s gate-before-train rule is the key design choice. By testing feasibility on the target accelerator and projecting cost before the main run, CAPO tries to prevent the common sequence in which a model begins training, exhausts budget, and returns no usable artifact. The preprint presents this as a form of compute admission control: an automated filter that decides not only how to adapt a model, but whether adaptation should happen at all.
That framing matters in biology, where model selection is increasingly constrained by compute rather than by theory alone. A workflow that can refuse an unaffordable path is not just conservative; it is operationally realistic.
Three Protein Tasks Show Where the Gate Helps
CAPO was evaluated without system changes across three protein-modeling tasks: SARS-CoV-2 host-range prediction, hit-to-lead affinity ranking, and temporal variant-effect reclassification. The benchmarks span different failure modes. One asks a model to resolve host specificity, another asks it to rank assay-linked binding outcomes, and the third tests whether a model can remain useful when variant interpretation moves forward in time.
On SARS-CoV-2 host-range prediction, CAPO partially fine-tuned an 8-million-parameter ESM-2 model and reached a species-averaged MCC of 0.826. That compared with 0.743 for a hand-designed expert transformer. CAPO also matched a general coding agent while using 57.4% less GPU time and reduced failed training trials from four to zero. The result is less about a single score and more about what the score cost to obtain: less compute, fewer dead ends, and a cleaner training history.
On hit-to-lead affinity ranking, CAPO rejected Boltz-2 fine-tuning altogether and chose a different path. Instead of burning cycles on a model adaptation with poor compute economics, it trained an assay-calibrated lightweight head over frozen representations. That strategy improved ranking on 11 of 15 held-out targets and completed in 1.68 GPU-hours. The biological logic here is straightforward: when the backbone already encodes useful sequence features, a smaller calibrated head can be enough to re-rank compounds or variants against a narrow assay without paying for full model surgery.
On temporally held-out ClinVar variants, rank-4 low-rank adaptation of ESM-2 reached an MCC of 0.478. The one-hot baseline achieved 0.252, and the frozen zero-shot score was 0.000. Time-aware evaluation is unforgiving, and the jump from zero-shot failure to modest but real performance suggests that parameter-efficient adaptation can recover clinically relevant signal when the data distribution has shifted.
Why Selective Computation Fits Protein Science
Protein-model adaptation is rarely just a machine-learning problem. It is also an allocation problem. Different assays tolerate different latencies, budgets, and levels of model drift, and the wrong choice can waste both compute and experimental attention. CAPO’s architecture answers that reality with a triage mindset: identify the task, estimate the cost, test the pathway, and proceed only when the expected return justifies the spend.
There is also a broader methodological implication. Many agentic systems treat fine-tuning as a dependable tool call, then infer feasibility after the fact. CAPO inverts that sequence. It asks whether the training plan is admissible before the job begins, which is closer to how a careful bench scientist would think about scarce reagents or limited instrument time. In computational terms, the gate protects the budget from avoidable failure modes; in biological terms, it preserves attention for the adaptations that are most likely to matter.
The audit trail is part of the value proposition as well. When a system can document why it trained, why it declined a path, and what it used instead, the workflow becomes easier to inspect and easier to reproduce. For model adaptation in biology, where downstream users may need to justify both cost and performance, that traceability is not ornamental.
Single-Run Evidence, Real Constraints, and What Still Needs Proof
The results are encouraging, but the evidentiary base remains narrow. These findings come from single runs, without a gate ablation and with only one generalist-agent comparison. Those limits matter because the strongest claim here is not that CAPO always wins, but that measured gating can change the economics of adaptation in meaningful ways.
One sentence captures the whole preprint: CAPO did not simply automate training, it decided when training should not happen.
Future validation would need to show whether the same admission-control logic holds across more protein families, more hardware configurations, and more uneven data regimes. It would also need to test whether the cost savings persist when the benchmarks become larger, noisier, or more clinically proximate. For now, the preprint shows a workable template for a smarter protein-model pipeline: measure first, train second, and choose the cheaper path when the biology does not justify the expense.